[Paper Review] Deep learning-based assessment of tumor-associated stroma for diagnosing breast cancer in histopathology images

Babak Ehteshami Bejnordi, Jimmy Linz+7 more|arXiv (Cornell University)|Feb 19, 2017

AI in cancer detection4 citations

TL;DR

This study proposes a deep learning system that diagnoses breast cancer by analyzing tumor-associated stroma in H&E-stained whole-slide images, using two convolutional neural networks (CNNs) to classify tissue components and stromal features, achieving an AUC of 0.921 in distinguishing cancer from benign disease—demonstrating stroma's strong discriminative power as a diagnostic biomarker.

ABSTRACT

Diagnosis of breast carcinomas has so far been limited to the morphological interpretation of epithelial cells and the assessment of epithelial tissue architecture. Consequently, most of the automated systems have focused on characterizing the epithelial regions of the breast to detect cancer. In this paper, we propose a system for classification of hematoxylin and eosin (H&E) stained breast specimens based on convolutional neural networks that primarily targets the assessment of tumor-associated stroma to diagnose breast cancer patients. We evaluate the performance of our proposed system using a large cohort containing 646 breast tissue biopsies. Our evaluations show that the proposed system achieves an area under ROC of 0.92, demonstrating the discriminative power of previously neglected tumor-associated stroma as a diagnostic biomarker.

Motivation & Objective

• To develop a data-driven system that diagnoses breast cancer by focusing on stromal morphology rather than epithelial features.
• To investigate whether tumor-associated stroma, previously overlooked, can serve as a robust diagnostic biomarker in histopathology.
• To overcome limitations of hand-crafted features by using deep learning to automatically learn stromal morphological representations.
• To enable whole-slide image classification at scale using end-to-end deep learning with minimal manual annotation.
• To evaluate the performance of stroma-based classification in distinguishing invasive breast cancer from benign breast disease.

Proposed method

• A VGG-like 11-layer CNN (CNN I) was trained to classify tissue into epithelium, stroma, and fat using whole-slide images.
• A second CNN (CNN II) was trained to classify stromal regions as normal or tumor-associated stroma using annotated regions from pathologists.
• 1×1 convolutional layers replaced fully connected layers to allow processing of arbitrary-sized whole-slide images.
• Features were extracted from both CNNs: global tissue distribution from CNN I and stromal characteristics from CNN II.
• A random forest classifier was trained on the combined feature set to classify whole slides as cancer or non-cancer.
• Hard negative mining was applied during training to improve generalization by retraining on false positives.

Experimental results

Research questions

• RQ1 Can tumor-associated stroma be used as a reliable diagnostic biomarker for breast cancer in H&E-stained whole-slide images?
• RQ2 Does a deep learning-based system that focuses on stromal morphology outperform epithelial-centric approaches in breast cancer classification?
• RQ3 Can end-to-end deep learning models automatically learn discriminative stromal features without relying on hand-crafted morphological descriptors?
• RQ4 What is the diagnostic performance of a stroma-focused system in distinguishing invasive breast cancer from benign breast disease?
• RQ5 How do stromal features compare to epithelial features in terms of predictive power for breast cancer diagnosis?

Key findings

• The proposed system achieved an area under the ROC curve (AUC) of 0.921 (95% CI: 0.862–0.967) at the whole-slide level, demonstrating high diagnostic accuracy.
• CNN I achieved 95.5% pixel-level accuracy in classifying tissue into epithelium, stroma, and fat.
• CNN II achieved 92.0% pixel-level accuracy in distinguishing normal stroma from tumor-associated stroma.
• The system's performance improved when combining features from both CNNs, with AUC of 0.921 compared to 0.882 (CNN I only) and 0.904 (CNN II only).
• The results confirm that stromal features alone contain sufficient discriminative information to diagnose breast cancer with high confidence.
• The study demonstrates the feasibility and potential of tumor-associated stroma as a novel, data-driven diagnostic biomarker in digital pathology.